Can AI predict future lung cancer risk from a chest CT?

How AI predicts lung cancer risk 6 years in advance

MIT Jameel Clinic - AI & Health · 2025-01-22 · Source

This English-language video, published by MIT Jameel Clinic on YouTube on 22 January 2025, explains Sybil’s research approach. It should be read alongside subgroup limitations from the later 2025 external test; it is not personal screening advice or proof of clinical benefit.

What question did the Sybil study ask?

The original study, published online on 12 January 2023, investigated future lung cancer risk prediction from the volume of a single low-dose CT. The model was developed using NLST data and evaluated on held-out NLST test data and data from Massachusetts General Hospital and Chang Gung Memorial Hospital in Taiwan. Retrospective image evaluation is distinct from an outcome study in which the model directs clinical decisions.

What did the 2025 external test add?

The Radiology study published on 5 August 2025 retrospectively examined data from 18,057 people at a single medical checkup facility. Cancers already visible on baseline CT were evaluated separately from cancers arising later. In the never- or light-smoking subgroup, the AUC for future cancers was 0.56. AUC measures discrimination; it does not mean a person has a 56% cancer risk or that diagnostic accuracy is 56%.

Why separate CT screening benefit from AI benefit?

The 2011 NLST publication reported a randomized trial comparing low-dose CT screening with chest radiography in people at high risk and found a reduction in lung cancer mortality. That result concerns a particular screening approach; it did not test the incremental benefit of using Sybil. Presenting screening benefit as a risk model’s performance confuses different types of evidence.

Does a risk score replace clinical decisions?

A risk estimate is neither a diagnosis demonstrating cancer in tissue nor a safe follow-up interval for everyone. The reviewed AI studies are not randomized implementation trials showing that changing follow-up intervals with the model improves patient outcomes. Publication of a paper, public availability of a model and regulatory authorization for a specific clinical use are also different states.

How should an implementation claim be assessed?

As an editorial appraisal, examine the target population, smoking history, CT protocol, model version and prediction horizon together. Subgroup performance, agreement between predicted and observed risk, false alarms and false reassurance need separate assessment. The next step is prospective testing of benefits and harms within actual decision workflows. This article gives no personal screening, follow-up or treatment advice.